{"id":"539e730e-50b9-4fbd-bd50-1c233f8f2e75","arxiv_id":"2411.09140","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A teacher-student semi-supervised model with uncertainty-weighted consistency and adversarial feature alignment segments ROP retinal vessels using unlabeled infant images and public labeled data.","lead":"The authors propose a semi-supervised framework that trains a teacher-student segmentation model on public labeled retinal images plus unlabeled retinopathy of prematurity (ROP) images, using uncertainty-weighted consistency and adversarial feature alignment. The method is evaluated against six baselines and is also used to produce vessel masks for ROP stage classification.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Vessel-unveiling target yw is under-specified: Eq. 4 substitutes yw for the teacher output, and the Fig. 3 softmax could produce a degenerate near-zero target, so the reported ROP gains rest on an unvalidated loss definition.","rationale":"The paper combines known SSL components; the authors should be credited for framing an important problem and for reporting baseline comparisons. The central claim, however, depends on the vessel-unveiling module, whose mathematical specification is not self-consistent. The reader's weakest_assumption focused on yw being a reliable target; I agree with that worry and add a more concrete failure mode: the definition of yw itself is ambiguous and potentially degenerate. This supports a CONDITIONAL verdict—the method may work, but it must be re-specified and validated before the reported margins can be trusted. I am not alleging any misrepresentation; a typo in Eq. 4 or an omitted axis in the softmax could easily explain the gap. The proposed check would distinguish a harmless typo from a substantive error, and the reader's existing CONDITIONAL verdict remains the right recommendation.","tokens_in":15401,"tokens_out":4002,"duration_ms":42129,"concrete_test":"Reimplement Section III-B exactly: compute p_k with MC dropout K=8, entropy = −Σ p log p, then yw = softmax(1 − entropy) * mean(p_k), with softmax over the full H×W map and also with a per-pixel sigmoid alternative. Train the proposed model with each variant on CHASEDB labeled + ROP unlabeled, then evaluate on the 12 ROP test annotations. If the exact-as-written variant yields near-zero yw (e.g., max value < 1e-3) or a Dice score within noise of the Mean Teacher baseline, while the sigmoid variant recovers the reported 45.24% Dice, then the paper's central module is mis-specified and the headline result depends on an undocumented deviation. Also recompute the loss with Eq. 4 corrected to use ˆy′_j; if results change materially, the loss term was implemented differently from the text.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section III-B defines the vessel-unveiling target yw, and the consistency loss Lcons in Eq. 3 contains two terms: α·MSE(ˆyj, ˆy′_j) + DIST(ˆyj, yw). The text says MSE is between student output and teacher output, but Eq. 4, which is labeled MSE(ˆyj, ˆy′_j), is written as (1/(H×W))Σ(ˆyj − yw)^2. Thus the equation contradicts the notation: the MSE term uses yw, not the teacher output. If the implementation follows Eq. 4 literally, the teacher output only enters through yw and the intended MSE-to-teacher term is missing. More seriously, the construction of yw in Fig. 3 is `softmax(1 − vessel_entropy) ⊗ pixel-wise average`. If softmax is applied over the full H×W entropy map (as written; the figure says `softmax(1−vessel_entropy)` with no axis specified), the weights sum to 1 and, when multiplied by a probability map, produce values on the order of 1/(H×W). For 400×400 patches, yw would be near zero everywhere, making Lcons degenerate and effectively pushing the student to predict all-background. The paper reports that this module improves ROP Dice from 39.00% (UCC) to 45.24%; if yw as written is near zero, that gain cannot come from the described loss. The module is the paper's central novelty, and the ROP test set is only 12 images, so the headline result rests on an unvalidated and internally inconsistent loss definition.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a semi-supervised retinal vessel segmentation framework for retinopathy of prematurity (ROP), combining a teacher-student consistency model with an uncertainty-weighted vessel-unveiling module and a feature-level adversarial domain alignment discriminator. The model is trained on publicly labeled datasets (CHASEDB, STARE) together with unlabeled in-house ROP images, and is evaluated on small ROP and public test sets. The authors report consistent improvements over several semi-supervised baselines, and they extend the framework to a downstream ROP multi-stage classification task in which predicted vessel masks are fused with fundus images.","tokens_in":15673,"tokens_out":4680,"duration_ms":132276,"significance":"If the reported results hold, the paper addresses a real annotation bottleneck in ROP vessel segmentation and demonstrates a potentially useful downstream diagnostic application. The work is among the first to attempt vessel segmentation across all four ROP stages without ground-truth ROP vessel annotations, and the combination of adversarial feature alignment with uncertainty-weighted teacher-student learning is a reasonable direction. The paper also reports results on multiple public datasets and a downstream classification task. However, the central vessel-unveiling loss is described inconsistently in the equations, the target construction is under-specified and potentially degenerate, and the evaluation is based on very small test sets without statistical significance assessment. These issues currently prevent verification of the central claims.","major_comments":[{"comment":"Equation (4) is labeled MSE(ŷ_j, ŷ'_j), but its right-hand side is (1/(H×W)) Σ (ŷ_j − yw)², using yw instead of ŷ'_j. With Equation (5), both consistency terms in Equation (3) compare the student output ŷ_j to the vessel-unveiling target yw, so the teacher output ŷ'_j never actually enters the consistency loss as written. If the implementation follows Eq. (4) literally, the intended teacher-student MSE term is absent and the loss reduces to two weighted variants of the same student-to-yw comparison. The authors must either correct the equation or clarify the implementation and confirm that the reported results are produced by the described loss.","section":"Section III-B, Eqs. (3)-(5)"},{"comment":"The construction of the vessel-unveiling target yw is under-specified and potentially degenerate. The operation softmax(1 − vessel_entropy) is written without specifying the axis of the softmax. If softmax is applied over the full H×W spatial map, the weights sum to 1 and, when multiplied elementwise by the averaged prediction, yield values on the order of 1/(H×W); for the 400×400 patches used in training this makes yw nearly zero everywhere, which would collapse the consistency loss toward predicting all-background. If softmax is applied over the channel dimension, the single-channel softmax is identically 1 and the uncertainty weighting has no effect. The authors must specify the intended normalization and provide evidence, e.g., histograms of yw on unlabeled ROP images, that the target is non-degenerate.","section":"Section III-B and Fig. 3"},{"comment":"The evaluation is severely underpowered for the central claims. The ROP test set consists of only 12 manually annotated images, and the public test sets consist of 5 images each. No standard deviations, confidence intervals, or significance tests are reported for any table. The headline result that the proposed model beats the second-best method by more than 6% Dice on the ROP test set (45.24% versus 39.00%) relies on a single split of 12 images. Please report repeated-seed experiments with variance measures and, if possible, a larger annotated ROP test set or a significance test (e.g., paired bootstrap over images).","section":"Section IV-B, IV-E, Tables I-III"},{"comment":"The downstream classification experiment uses 3,873 fundus images from 217 patients, with approximately 5 images per eye, but the paper does not state how the train/test split was performed. If images from the same patient or the same eye appear in both training and test partitions, the reported fusion improvement (76.38% versus 74.76% accuracy) will be inflated by intra-patient correlation. The authors should specify a patient-level or eye-level split and report per-patient or per-eye metrics, together with confidence intervals, before claiming clinical utility for the downstream task.","section":"Section V-A and V-C"}],"minor_comments":[{"comment":"The sentence 'In our work, we make four three contributions' contains an editing error; it should list either four or three contributions consistently.","section":"Section I, Contributions"},{"comment":"The vessel-entropy formula in Figure 3 is written as '−1.0×K ∑ipilog(pi)', which is ambiguous. It should clarify whether the average is over the K dropout passes and whether the entropy is computed over the class dimension of the probabilities.","section":"Fig. 3"},{"comment":"The variable Ivessel is called 'vessel entropy' but it is not defined whether higher Ivessel corresponds to higher uncertainty or higher vessel probability. Since Eq. (5) multiplies the distance by (1−Ivessel), the meaning of Ivessel is load-bearing for interpreting the weighting.","section":"Eq. 5"},{"comment":"The conclusion refers to a 'vessel-veiling module' where the rest of the paper uses 'vessel-unveiling module'; this typo should be corrected.","section":"Section VI, Conclusion"},{"comment":"The captions state 'UpperBound' is fully supervised with only the labeled public images. This is clear, but the tables would benefit from a note that no ROP ground-truth vessels are used in training, to avoid confusion with the 12 held-out ROP test annotations.","section":"Tables I-III"}],"recommendation":"major_revision","confidential_remarks":"The contribution hinges on the vessel-unveiling consistency loss, but the mismatch between Eq. (3) and Eq. (4) plus the unspecified softmax axis make it currently impossible to verify that the reported ROP Dice of 45.24% is produced by the described mechanism. This is fixable in principle if the equations were typos and the implementation is non-degenerate, which is why I recommend major revision rather than rejection. The small test sets and the absence of a patient-level split in the downstream task are additional risks that should be addressed in revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague—\n\nHere's my take on the ROP vessel segmentation paper (arXiv:2411.09140). The application is genuinely new: using semi-supervised learning to segment infant ROP vessels from public labeled datasets plus unlabeled ROP images, then feeding those masks into a stage classifier. The authors compare against six baselines on five metrics, run ablations, and report a downstream classification task. That's solid execution of a sensible pipeline.\n\nThe public-dataset results (CHASEDB, STARE) are competitive and consistent across label-unlabeled settings; the ablation shows monotonic gains from each added component. The classification fusion experiment is a useful addition, even if the accuracy improvement over fundus-only is modest (~1.6 points). The paper is clearly written and the method is described well enough to reproduce in principle.\n\nThe soft spots are real, and they matter for the headline claims. First, Eq. 4 defines MSE(ˆyj, ˆy′_j) using yw in the RHS, which contradicts the text and makes the two consistency terms in Eq. 3 nearly redundant. That reads as a typo, but it's exactly the kind of inconsistency that needs fixing before the loss definition can be trusted. Second, Fig. 3's softmax axis is unspecified; if softmax is taken over the full spatial map, yw would be near zero, which would push the student to predict all background and couldn't produce the reported 6-point Dice gain. So either the implementation is different from the description, or the description is missing a crucial detail. Third, the ROP test set has only 12 manually annotated images, public test sets have 5, and no error bars or significance tests are reported. The downstream classification gain is small, and clinical significance is plausible but unproven. Code and data aren't released, which limits reproducibility.\n\nThe circularity burden is low: the held-out annotations are external, and the train/test split is described as separate. But the vessel-unveiling target yw is unvalidated as a training signal—there's no evidence that the uncertainty-weighted targets are accurate before training, and if they emphasize false positives, the consistency loss could reinforce them.\n\nThis paper deserves a serious referee. The application is worthwhile, and the public-dataset results are likely to be roughly correct after correction. I'd send it to review with a request to clarify the vessel-unveiling module, fix Eq. 4, and report variance or use a larger ROP test set. I'd cite it only after those issues are addressed.","headline":"A credible SSL pipeline for ROP vessel segmentation with solid public-dataset results, but a key loss equation and softmax axis are under-specified and the ROP test set is thin.","tokens_in":16314,"tokens_out":3718,"would_cite":false,"duration_ms":34481,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A semi-supervised teacher–student method segments retinal vessels in retinopathy of prematurity (ROP) without manual ROP annotations, and its vessel masks improve four-stage ROP classification when fused with fundus images.","keywords":["semi-supervised segmentation","retinal vessel segmentation","retinopathy of prematurity","teacher-student learning","domain adversarial learning","uncertainty-weighted consistency","multi-stage ROP classification"],"falsifier":"Compare the vessel-unveiling target $y_w$ with the 12 manually annotated ROP test images: if $y_w$ has a lower Dice against ground truth than the teacher's plain averaged prediction, then the entropy weighting is selecting noise rather than vessels, and the consistency loss is training the student toward a worse target.","tokens_in":15113,"feed_emoji":"🩺","tokens_out":8679,"duration_ms":86812,"temperature":0.7,"pith_summary":"Retinopathy of prematurity (ROP) is a leading cause of blindness in preterm infants, and its diagnosis depends on reading retinal vessel structure; but preterm vessels are thin and faint, and manual annotation is prohibitively difficult. This paper argues that a semi-supervised teacher–student segmentation model can extract those vessels without any manually annotated ROP image, using only 18 labeled adult/child retinal images from public datasets plus 90 unlabeled ROP images. The two technical proposals are an uncertainty-weighted vessel-unveiling module, which amplifies faint or hidden vessels in the teacher's predictions, and a domain-adversarial discriminator that aligns feature distributions across the public and ROP domains. If the argument holds, clinicians could get vessel maps for ROP staging at no annotation cost, and the vessel maps themselves improve multi-stage classification accuracy when fused with original fundus images. The central evidence is a Dice score of 45.24% on the ROP test set, over 6 percentage points above the second-best baseline.","feed_headline":"Vessel maps for preterm eyes arrive without manual ROP labels","feed_subtitle":"Fusing the extracted vessel masks with fundus images lifts ROP stage classification from 74.76% to 76.38%.","key_machinery":"The load-bearing mechanism is the uncertainty-weighted vessel-unveiling module inside the teacher branch. For each unlabeled ROP image, the teacher performs $K=8$ stochastic forward passes with dropout and soft augmentation, producing probability maps $p_k$; a per-pixel vessel entropy $I_{\\mathrm{vessel}}$ is derived across these passes, and the softmax of $(1 - I_{\\mathrm{vessel}})$ is element-wise multiplied with the averaged prediction to form the vessel-unveiling target $y_w$. The consistency loss combines an unweighted term comparing the student to the teacher's averaged prediction and a distance to $y_w$ weighted by $(1 - I_{\\mathrm{vessel}})$, so the student is pushed toward faint vessels while being allowed to stay uncertain where the teacher is uncertain. A PatchGAN-style discriminator takes encoder features from labeled and unlabeled images and adversarially aligns the public and ROP domains, and the student uses three regularized decoders (main, feature-noise, feature-dropout) with supervised loss on the labeled public images.","core_discovery":"On its own terms, the paper claims that combining adversarial feature alignment with an uncertainty-weighted consistency target enables a teacher–student network to segment retinal vessels in unlabeled ROP fundus images, despite a large domain gap from the public datasets it trains on. The method is the first to attempt ROP vessel segmentation in a semi-supervised setting without any ROP ground-truth vessel annotations during training; the only ROP annotations are 12 manually traced test images used for evaluation. The reported Dice of 45.24% on the ROP test set, exceeding the second-best baseline by more than 6%, is taken as evidence that the vessel-unveiling module and domain alignment extract vessels that other semi-supervised methods miss. The paper further shows that feeding the segmented vessel masks into a fusion classifier with the original fundus images raises four-stage ROP classification accuracy from 74.76% to 76.38%.","pith_inferences":["Editorial inference: the paper never measures overlap between the entropy-weighted target $y_w$ and true ROP vessels; computing that overlap on the 12 annotated test images would settle whether the unveiling module selects vessels or noise.","Editorial inference: the reported ROP gain might come from the teacher's averaged prediction rather than the entropy weighting; a control with a plain averaged teacher target would isolate the module's contribution.","Editorial inference: if the domain-adversarial alignment transfers public-dataset knowledge to ROP images, the same recipe can be tested on other pediatric imaging domains with scarce infant annotations.","Editorial inference: the 1.62-point accuracy gain from fusing vessel masks could reflect extra model capacity; a control with a second image channel carrying no vessel information would test that."],"forward_implications":["ROP vessel segmentation can be performed without any manually annotated ROP images, removing the main annotation bottleneck for this disease.","The model retains source-domain performance on CHASEDB and STARE while improving ROP test Dice, showing the adversarial alignment does not sacrifice labeled-domain accuracy.","Vessel masks from the segmentation model carry complementary signal: fusing them with fundus images raises four-stage ROP classification accuracy from 74.76% to 76.38% and precision from 71.61% to 75.79%.","The same semi-supervised recipe can be applied to other retinal vessel segmentation tasks where target-domain labels are unavailable."],"supporting_citations":[{"why":"Mean Teacher: source of the EMA teacher-student consistency framework the method extends.","marker":"[22]"},{"why":"Uncertainty-aware self-ensembling: origin of using uncertainty maps from multiple forward passes to guide consistency training.","marker":"[18]"},{"why":"MC dropout: supplies the K stochastic forward passes used to estimate per-pixel vessel entropy.","marker":"[16]"},{"why":"PatchGAN: design basis for the discriminator that produces matrix outputs for adversarial feature alignment.","marker":"[23]"},{"why":"CHASEDB: the main labeled source domain dataset (18 images) for training.","marker":"[3]"},{"why":"STARE: additional public labeled dataset used to test cross-domain generalization.","marker":"[24]"},{"why":"U-Net: the architecture of both student and teacher segmentation networks.","marker":"[19]"},{"why":"Domain-adversarial neural networks: the principle of gradient-based feature alignment that the discriminator realizes.","marker":"[42]"}],"fun_headline_variants":["Semi-supervised vessel unveiling for preterm eye diagnosis","Adversarial learning segments preterm vessels with no labels","Retinal vessel maps for ROP without manual tracing","Teacher-student network reveals vessels in preterm eyes","Domain alignment lifts ROP vessel segmentation from unlabeled data"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The teacher's uncertainty-weighted vessel-unveiling prediction $y_w$ is adopted as a reliable learning target for unlabeled ROP images, but it is never checked against any annotated ROP vessel map before training starts; if the uncertainty weighting emphasizes false positives, the consistency loss will amplify them and the reported ROP gains will not transfer.","fun_headline_variants_meta":{"raw":{"variants":["Semi-supervised vessel unveiling for preterm eye diagnosis","Adversarial learning segments preterm vessels with no labels","Retinal vessel maps for ROP without manual tracing","Teacher-student network reveals vessels in preterm eyes","Domain alignment lifts ROP vessel segmentation from unlabeled data"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000343,"raw_usage":{"total_tokens":1907,"prompt_tokens":985,"completion_tokens":922,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":601,"completion_tokens_details":{"reasoning_tokens":845}},"tokens_in":601,"tokens_out":922,"duration_ms":32906,"temperature":1.0,"reasoning_tokens":845,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T20:59:16.732325+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compare the vessel-unveiling target $y_w$ with the 12 manually annotated ROP test images: if $y_w$ has a lower Dice against ground truth than the teacher's plain averaged prediction, then the entropy weighting is selecting noise rather than vessels, and the consistency loss is training the student toward a worse target.","supporting_citations":[{"cited_title":"Ronneberger, P","cited_arxiv_id":null,"evidence_quote":"U-Net: the architecture of both student and teacher segmentation networks."},{"cited_title":"Yaroslav, E","cited_arxiv_id":null,"evidence_quote":"Domain-adversarial neural networks: the principle of gradient-based feature alignment that the discriminator realizes."},{"cited_title":"Tarvainen and H","cited_arxiv_id":null,"evidence_quote":"Mean Teacher: source of the EMA teacher-student consistency framework the method extends."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Uncertainty-aware self-ensembling: origin of using uncertainty maps from multiple forward passes to guide consistency training."},{"cited_title":"Gal and Z","cited_arxiv_id":null,"evidence_quote":"MC dropout: supplies the K stochastic forward passes used to estimate per-pixel vessel entropy."},{"cited_title":"Isola, J","cited_arxiv_id":null,"evidence_quote":"PatchGAN: design basis for the discriminator that produces matrix outputs for adversarial feature alignment."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"CHASEDB: the main labeled source domain dataset (18 images) for training."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"STARE: additional public labeled dataset used to test cross-domain generalization."}],"review_version":1}